ERα靶向化合物筛选与优化：基于深度学习和多目标优化的ADMET模型

ERα Targeted Compound Screening and Optimization: ADMET Model Based on Deep Learning and Multi-Objective Optimization

摘要: 目前，抗胰腺癌候选药物化合物在药物研发中面临时间和成本等诸多挑战。因此，本文提出一种融合Lasso回归与BP神经网络模型的方法，用于筛选和优化ERα靶向化合物。首先，使用Lasso回归筛选出与生物活性(pIC50)相关的重要分子描述符，并通过神经网络进行ADMET分类预测。实验结果表明，该方法能够有效提高药物活性和安全性的预测精度。从Lasso回归中筛选出的前20个重要特征对药物活性有显著影响，构建的随机森林回归模型在测试集上的准确率达到89%。并且筛选的特征在BP神经网络中ADMET分类任务中也表现良好，其中CYP3A4任务的准确率为91%。该方法为ERα拮抗剂的筛选和优化提供了可借鉴的思路。

Abstract: Currently, anti-breast cancer drug candidate compounds are facing many heavy challenges in drug discovery such as time and cost. Therefore, in this paper, we propose an approach that integrates Lasso regression and BP neural network models for screening and optimizing ERα-targeting compounds. First, important molecular descriptors related to biological activity (pIC50) were screened using Lasso regression and predicted by neural network for ADMET classification. The experimental results showed that this method can effectively improve the prediction accuracy of drug activity and safety. The top 20 important features screened from Lasso regression had a significant effect on drug activity, and the accuracy of the constructed random forest regression model reached 89% on the test set. And the screened features also performed well in the ADMET classification task in BP neural network, with an accuracy of 91% in the CYP3A4 task. This method provides a referable idea for the screening and optimization of ERα antagonists.